Humans are harder to scale, and it took billions of years for evolution to get here, with enormous resource and energy usage. A brain trained by evolution is already fit for the environment niche it has to inhabit. But an AI model has none of that, no evolution selecting the internal structure to be optimal. So they have to compensate by learning these things from tons of raw data. We are great at some tasks that relate to our survival, but bad at other tasks, even worse than other animals or AIs - we are not generally intelligent either.

Also, most AIs don't have real time interaction with the world. They only have restricted text interfaces or APIs, no robotic bodies, no way to do interventions to distinguish causal relations from correlations. When an AI has feedback loop from the environment it gets much better at solving tasks.